42 Background: Methods of stratifying esophageal adenocarcinoma patients into prognostic groups are needed, as are new insights into genetic determinants of disease behaviour. Prognosis is likely to have non-negligible genetic influences, as mediated by host responses to tumor, resistance to therapeutic side-effects, and/or an influence on tumor development. Prior studies have used candidate-gene approaches. We took an alternative approach, using an unbiased, genome-wide approach, and novel analytic methods that may be better able to detect multi-gene interactions, which may contribute the majority of genetic effects for many clinical phenotypes. Methods: Germline DNA from a Toronto-based cohort of EAC patients (n=270) was analyzed by Omni1 Quad microarray as part of the BEAGESS initiative. Quality control and analysis was performed using PLINK, R, and GenABEL software packages. A Cox proportional hazards (CPH) model for progression-free survival tested each polymorphism for independent effects at a genome-wide significance level of P < 1E-07, adjusting for population stratification. While classical analysis has limited ability to detect gene-gene interactions, a Random Survival Forest algorithm was used to detect effects based on the complex interactions among top 1,000 polymorphisms by p-value ranking. Results: After data cleaning and standard GWAS quality control procedures, there were 735,309 SNPs and 245 patients remaining for analysis. The CPH model, adjusted for population stratification, produced a satisfactory Q-Q plot, and showed one SNP (rs7844673, Chr 8) that was significant at p=7.8E-8. In addition, Random Forest based variable selection produced a set of 20 polymorphisms that (1) reproduced 86% of the predictive ability of the full 1000 variables, and (2) also included the #3 ranked polymorphism by CPH modeling (rs9290822, Chr 3) upstream of the IGF2BP2 gene. Conclusions: A genome-wide approach has discovered two previously undescribed SNPs with a potential influence on EAC prognosis via a combination of independent and interactive effects. Validation in an independent cohort is currently being pursued.
Imputation and quality control steps for combining multiple genome-wide datasets
